Derived facts are not rewound by as_of in entity fact queries

[ansonnmm]

Summary

When querying an entity with a record-time as_of, asserted facts are correctly rewound to the requested belief state, but derived facts are still returned from the current state.

This can cause a historical belief query to contain derivations that did not exist yet at the requested as_of time.

Tested version

Tested against stock Utopia at:

b03a0f20a3d053323a3ec2669c7c9b4aeb9a8e14

No source, schema, or migration modifications were made.

Reproduction

1. Create a knowledge base with entities and asserted facts that allow a rule/inference to produce a derived fact.
2. Record the timestamps before those entities/facts are ingested.
3. After ingestion and derivation complete, query the entity through MCP entity_facts using an as_of timestamp from before the entity/facts existed.
4. Compare asserted and derived results.

Actual behavior

The asserted facts are correctly absent at the earlier as_of.

However, a derived fact created later is still returned.

In my reproduction, a query at a record-time before the relevant entity/facts had been ingested still returned a later derivation similar to:

Atlas · part of · Acme Systems [rule: transitive]

So the effective result is:

asserted state: rewound by as_of
derived state: current

Expected behavior

A record-time query should represent one internally consistent belief snapshot.

If a derived fact did not exist, or its premises were not present, at the requested as_of, that derived fact should not appear.

Conceptually:

belief_as_of(T)
=
asserted_facts_as_of(T)
+
derived_facts_as_of(T)

rather than:

belief_as_of(T)
=
asserted_facts_as_of(T)
+
derived_facts_now

This matters especially for historical decision review or any other use case that asks:

What did this knowledge base believe at time T?

Relevant code path

At the tested commit, the asserted query receives as_of, while the derived query appears not to:

utopia/crates/utopia-server/src/api/tools.rs

Lines 307 to 370 in b03a0f2

	pub async fn entity_facts(ctx: &ToolCtx<'_>, args: &serde_json::Value) -> ToolResult {
	let id = args["entity_id"]
	.as_str()
	.and_then(|s| s.parse::<Uuid>().ok());
	// 世界轴过滤在 SQL 里（world_axis，0022）：没起点的事实从最早的证据起，结束了
	// 不知哪天的到说出它的那份文档为止。这里只把 T 传下去，不自己解释 NULL
	let at = args["at"].as_str().and_then(parse_when);
	// 记录轴（0019 / #347）：那一刻**我们持有**的事实。两根轴两个参数，绝不合成
	// 一个——合起来就会拿「三月的世界，以今天的认知」去答「三月的世界，以三月的认知」
	// 「更正到来之前」（#416）：`before` 是 changes 里印出来的那个时刻，原样抄过来。
	// 账本的钟是微秒，「严格早于 T」就是「不晚于 T 减一微秒」——这一步在这里做，
	// 不让模型对着 ISO 字符串算小数秒的借位：算错一位，答的就是更正**之后**的状态，
	// 而且看不出来（#351 那种错）。给了 before 就以它为准
	let before = args["before"].as_str().and_then(parse_when);
	let as_of = match before {
	Some(t) => Some(just_before(t)),
	None => args["as_of"].as_str().and_then(parse_when),
	};
	let Some(id) = id else {
	return (
	"Invalid entity_id (expected the uuid returned by find_entities).".to_string(),
	json!({ "kind": "facts", "label": "?", "detail": "invalid id" }),
	);
	};
	match utopia_store::graph::entity_detail(&ctx.state.pool, ctx.kb_id, id, at, as_of).await {
	Ok((node, facts)) => {
	// 规则的结论也是这个实体的一部分（0021）。**不给的话模型会拿那些
	// 读数自己再判一遍**——而阈值写在规则里，它看不见，于是两处判断
	// 迟早不一致，agent 那次还没有前提链、没有区间、也不进账本
	let derived =
	utopia_store::reasoning::derived_for_entity(&ctx.state.pool, ctx.kb_id, id, at)
	.await
	.unwrap_or_default();
	let mut derived: Vec<String> = derived
	.iter()
	.map(|d| {
	format!(
	"{} · {} · {} [rule: {}]",
	d.subject,
	d.predicate,
	d.object,
	d.rule_name.as_deref().unwrap_or(&d.rule),
	)
	})
	.collect();
	let text = if facts.is_empty() && derived.is_empty() {
	match at {
	Some(t) => format!(
	"{}: no facts valid as of {}.",
	node.name,
	crate::time_text::instant(t)
	),
	None => format!("{}: no recorded facts.", node.name),
	}
	} else {
	let mut lines: Vec<String> = facts.iter().map(fact_line).collect();
	lines.append(&mut derived);
	lines.join("\n")
	};
	let detail = entity_facts_detail(facts.len(), at, as_of, before);
	(
	text,
	json!({ "kind": "facts", "label": node.name, "detail": detail }),

The same semantic question may also apply to derived proof expansion: if a historical derived fact/proof is exposed, its premises and evidence should correspond to the same record-time snapshot.

Suggested acceptance criteria

I am not prescribing a particular implementation, but a regression test could establish the invariant:

1. Create asserted facts at record-time T2.
2. Materialize a derivation after T2.
3. Query with as_of=T1, where T1 < T2.
4. Neither the later asserted facts nor derivations based on them are returned.
5. Query with current as_of and both are returned.
6. If proof retrieval supports historical queries, the proof and its premises obey the same record-time semantics.

Why I think this is a correctness issue rather than a feature request

Utopia already exposes at and as_of as distinct world-time and record-time semantics. The unexpected behavior is that one category of graph state obeys the record-time axis while another category in the same response does not.

If this behavior is intentional, clarification of the intended semantics would also be useful.

[WaylandYang]

Confirmed, and it is a gap against our own record rather than an open semantic question. 0019 says derived_facts carries derived_at and invalidated_at so that derivations follow the same held_at rule as assertions, and the graph overview does that. The entity-level read, derived_for_entity, had kept the older invalidated_at IS NULL, and both the entity panel and entity_facts over MCP go through it. So a rewound entity showed March's assertions with today's conclusions, which is exactly the split you describe.

#557 gives that read as_of with the predicate the overview already uses, and pins both directions in a database test: a derivation is absent at a moment before it was drawn, and still present at a moment before it was withdrawn. Your acceptance criteria 1 to 5 are what the test does.

On 6: a proof has one version. fact_derivations is rewritten in place when a kept conclusion is reproved (0030), so expanding a rewound derivation shows today's premises, with the retracted ones marked. That is recorded as an open question in 0019 rather than changed here; the row's derived_at says when the conclusion first stood, and versioning the proofs is a larger change than anyone has asked for yet.

Thanks for the precise report and the pointer into tools.rs.